A gradient-based technique to add imperceptable Noise to an image that increases the loss function. Makes it impossible to train models off these images.